1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv/up-avx512.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <immintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f(
19     size_t channels,
20     size_t output_width,
21     const float** input,
22     const float* weights,
23     float* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const float* zero,
28     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
34   const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
35   do {
36     const float* i0 = input[0];
37     assert(i0 != NULL);
38     if XNN_UNPREDICTABLE(i0 != zero) {
39       i0 = (const float*) ((uintptr_t) i0 + input_offset);
40     }
41     const float* i1 = input[1];
42     assert(i1 != NULL);
43     if XNN_UNPREDICTABLE(i1 != zero) {
44       i1 = (const float*) ((uintptr_t) i1 + input_offset);
45     }
46     const float* i2 = input[2];
47     assert(i2 != NULL);
48     if XNN_UNPREDICTABLE(i2 != zero) {
49       i2 = (const float*) ((uintptr_t) i2 + input_offset);
50     }
51     const float* i3 = input[3];
52     assert(i3 != NULL);
53     if XNN_UNPREDICTABLE(i3 != zero) {
54       i3 = (const float*) ((uintptr_t) i3 + input_offset);
55     }
56     const float* i4 = input[4];
57     assert(i4 != NULL);
58     if XNN_UNPREDICTABLE(i4 != zero) {
59       i4 = (const float*) ((uintptr_t) i4 + input_offset);
60     }
61     const float* i5 = input[5];
62     assert(i5 != NULL);
63     if XNN_UNPREDICTABLE(i5 != zero) {
64       i5 = (const float*) ((uintptr_t) i5 + input_offset);
65     }
66     const float* i6 = input[6];
67     assert(i6 != NULL);
68     if XNN_UNPREDICTABLE(i6 != zero) {
69       i6 = (const float*) ((uintptr_t) i6 + input_offset);
70     }
71     const float* i7 = input[7];
72     assert(i7 != NULL);
73     if XNN_UNPREDICTABLE(i7 != zero) {
74       i7 = (const float*) ((uintptr_t) i7 + input_offset);
75     }
76     const float* i8 = input[8];
77     assert(i8 != NULL);
78     if XNN_UNPREDICTABLE(i8 != zero) {
79       i8 = (const float*) ((uintptr_t) i8 + input_offset);
80     }
81     const float* i9 = input[9];
82     assert(i9 != NULL);
83     if XNN_UNPREDICTABLE(i9 != zero) {
84       i9 = (const float*) ((uintptr_t) i9 + input_offset);
85     }
86     const float* i10 = input[10];
87     assert(i10 != NULL);
88     if XNN_UNPREDICTABLE(i10 != zero) {
89       i10 = (const float*) ((uintptr_t) i10 + input_offset);
90     }
91     const float* i11 = input[11];
92     assert(i11 != NULL);
93     if XNN_UNPREDICTABLE(i11 != zero) {
94       i11 = (const float*) ((uintptr_t) i11 + input_offset);
95     }
96     const float* i12 = input[12];
97     assert(i12 != NULL);
98     if XNN_UNPREDICTABLE(i12 != zero) {
99       i12 = (const float*) ((uintptr_t) i12 + input_offset);
100     }
101     const float* i13 = input[13];
102     assert(i13 != NULL);
103     if XNN_UNPREDICTABLE(i13 != zero) {
104       i13 = (const float*) ((uintptr_t) i13 + input_offset);
105     }
106     const float* i14 = input[14];
107     assert(i14 != NULL);
108     if XNN_UNPREDICTABLE(i14 != zero) {
109       i14 = (const float*) ((uintptr_t) i14 + input_offset);
110     }
111     const float* i15 = input[15];
112     assert(i15 != NULL);
113     if XNN_UNPREDICTABLE(i15 != zero) {
114       i15 = (const float*) ((uintptr_t) i15 + input_offset);
115     }
116     const float* i16 = input[16];
117     assert(i16 != NULL);
118     if XNN_UNPREDICTABLE(i16 != zero) {
119       i16 = (const float*) ((uintptr_t) i16 + input_offset);
120     }
121     const float* i17 = input[17];
122     assert(i17 != NULL);
123     if XNN_UNPREDICTABLE(i17 != zero) {
124       i17 = (const float*) ((uintptr_t) i17 + input_offset);
125     }
126     const float* i18 = input[18];
127     assert(i18 != NULL);
128     if XNN_UNPREDICTABLE(i18 != zero) {
129       i18 = (const float*) ((uintptr_t) i18 + input_offset);
130     }
131     const float* i19 = input[19];
132     assert(i19 != NULL);
133     if XNN_UNPREDICTABLE(i19 != zero) {
134       i19 = (const float*) ((uintptr_t) i19 + input_offset);
135     }
136     const float* i20 = input[20];
137     assert(i20 != NULL);
138     if XNN_UNPREDICTABLE(i20 != zero) {
139       i20 = (const float*) ((uintptr_t) i20 + input_offset);
140     }
141     const float* i21 = input[21];
142     assert(i21 != NULL);
143     if XNN_UNPREDICTABLE(i21 != zero) {
144       i21 = (const float*) ((uintptr_t) i21 + input_offset);
145     }
146     const float* i22 = input[22];
147     assert(i22 != NULL);
148     if XNN_UNPREDICTABLE(i22 != zero) {
149       i22 = (const float*) ((uintptr_t) i22 + input_offset);
150     }
151     const float* i23 = input[23];
152     assert(i23 != NULL);
153     if XNN_UNPREDICTABLE(i23 != zero) {
154       i23 = (const float*) ((uintptr_t) i23 + input_offset);
155     }
156     const float* i24 = input[24];
157     assert(i24 != NULL);
158     if XNN_UNPREDICTABLE(i24 != zero) {
159       i24 = (const float*) ((uintptr_t) i24 + input_offset);
160     }
161     input = (const float**) ((uintptr_t) input + input_stride);
162 
163     size_t c = channels;
164     const float* w = weights;
165     for (; c >= 32; c -= 32) {
166       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
167       __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
168 
169 
170       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
171       const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
172       i0 += 32;
173 
174       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
175       const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
176       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
177       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
178 
179       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
180       const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
181       i1 += 32;
182 
183       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
184       const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
185       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
186       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
187 
188       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
189       const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
190       i2 += 32;
191 
192       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
193       const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
194       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
195       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
196 
197       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
198       const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16);
199       i3 += 32;
200 
201       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
202       const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144);
203       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
204       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
205 
206       const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
207       const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16);
208       i4 += 32;
209 
210       const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
211       const __m512 vk4xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 176);
212       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
213       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
214 
215       const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
216       const __m512 vi5xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i5 + 16);
217       i5 += 32;
218 
219       const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
220       const __m512 vk5xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 208);
221       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
222       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
223 
224       const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
225       const __m512 vi6xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i6 + 16);
226       i6 += 32;
227 
228       const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
229       const __m512 vk6xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 240);
230       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
231       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
232 
233       const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
234       const __m512 vi7xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i7 + 16);
235       i7 += 32;
236 
237       const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
238       const __m512 vk7xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 272);
239       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
240       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
241 
242       const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
243       const __m512 vi8xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i8 + 16);
244       i8 += 32;
245 
246       const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
247       const __m512 vk8xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 304);
248       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
249       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
250 
251       const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
252       const __m512 vi9xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i9 + 16);
253       i9 += 32;
254 
255       const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
256       const __m512 vk9xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 336);
257       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
258       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
259 
260       const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
261       const __m512 vi10xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i10 + 16);
262       i10 += 32;
263 
264       const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
265       const __m512 vk10xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 368);
266       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
267       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
268 
269       const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
270       const __m512 vi11xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i11 + 16);
271       i11 += 32;
272 
273       const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
274       const __m512 vk11xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 400);
275       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
276       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
277 
278       const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
279       const __m512 vi12xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i12 + 16);
280       i12 += 32;
281 
282       const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
283       const __m512 vk12xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 432);
284       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
285       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
286 
287       const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
288       const __m512 vi13xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i13 + 16);
289       i13 += 32;
290 
291       const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
292       const __m512 vk13xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 464);
293       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
294       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
295 
296       const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
297       const __m512 vi14xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i14 + 16);
298       i14 += 32;
299 
300       const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
301       const __m512 vk14xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 496);
302       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
303       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
304 
305       const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
306       const __m512 vi15xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i15 + 16);
307       i15 += 32;
308 
309       const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
310       const __m512 vk15xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 528);
311       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
312       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
313 
314       const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
315       const __m512 vi16xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i16 + 16);
316       i16 += 32;
317 
318       const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
319       const __m512 vk16xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 560);
320       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
321       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
322 
323       const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
324       const __m512 vi17xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i17 + 16);
325       i17 += 32;
326 
327       const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
328       const __m512 vk17xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 592);
329       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
330       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
331 
332       const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
333       const __m512 vi18xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i18 + 16);
334       i18 += 32;
335 
336       const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
337       const __m512 vk18xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 624);
338       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
339       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
340 
341       const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
342       const __m512 vi19xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i19 + 16);
343       i19 += 32;
344 
345       const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
346       const __m512 vk19xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 656);
347       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
348       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
349 
350       const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
351       const __m512 vi20xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i20 + 16);
352       i20 += 32;
353 
354       const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
355       const __m512 vk20xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 688);
356       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
357       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
358 
359       const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
360       const __m512 vi21xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i21 + 16);
361       i21 += 32;
362 
363       const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
364       const __m512 vk21xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 720);
365       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
366       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
367 
368       const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
369       const __m512 vi22xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i22 + 16);
370       i22 += 32;
371 
372       const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
373       const __m512 vk22xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 752);
374       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
375       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
376 
377       const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
378       const __m512 vi23xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i23 + 16);
379       i23 += 32;
380 
381       const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
382       const __m512 vk23xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 784);
383       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
384       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
385 
386       const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
387       const __m512 vi24xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i24 + 16);
388       i24 += 32;
389 
390       const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
391       const __m512 vk24xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 816);
392       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
393       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
394 
395       w += 832;
396 
397 
398       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
399       __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin);
400       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
401       vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax);
402 
403       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
404       _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV);
405       output += 32;
406     }
407     for (; c >= 16; c -= 16) {
408       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
409 
410       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
411       i0 += 16;
412 
413       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
414       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
415 
416       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
417       i1 += 16;
418 
419       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
420       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
421 
422       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
423       i2 += 16;
424 
425       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
426       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
427 
428       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
429       i3 += 16;
430 
431       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
432       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
433 
434       const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
435       i4 += 16;
436 
437       const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
438       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
439 
440       const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
441       i5 += 16;
442 
443       const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
444       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
445 
446       const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
447       i6 += 16;
448 
449       const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
450       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
451 
452       const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
453       i7 += 16;
454 
455       const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
456       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
457 
458       const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
459       i8 += 16;
460 
461       const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
462       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
463 
464       const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
465       i9 += 16;
466 
467       const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
468       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
469 
470       const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
471       i10 += 16;
472 
473       const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
474       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
475 
476       const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
477       i11 += 16;
478 
479       const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
480       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
481 
482       const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
483       i12 += 16;
484 
485       const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
486       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
487 
488       const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
489       i13 += 16;
490 
491       const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
492       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
493 
494       const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
495       i14 += 16;
496 
497       const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
498       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
499 
500       const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
501       i15 += 16;
502 
503       const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
504       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
505 
506       const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
507       i16 += 16;
508 
509       const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
510       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
511 
512       const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
513       i17 += 16;
514 
515       const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
516       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
517 
518       const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
519       i18 += 16;
520 
521       const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
522       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
523 
524       const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
525       i19 += 16;
526 
527       const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
528       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
529 
530       const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
531       i20 += 16;
532 
533       const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
534       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
535 
536       const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
537       i21 += 16;
538 
539       const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
540       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
541 
542       const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
543       i22 += 16;
544 
545       const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
546       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
547 
548       const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
549       i23 += 16;
550 
551       const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
552       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
553 
554       const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
555       i24 += 16;
556 
557       const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
558       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
559 
560       w += 16;
561 
562 
563       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
564       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
565 
566       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
567       output += 16;
568     }
569     if XNN_UNLIKELY(c != 0) {
570       assert(c >= 1);
571       assert(c <= 16);
572       // Prepare mask for valid 32-bit elements (depends on nc).
573       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
574 
575       __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
576 
577       const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
578       const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
579       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
580 
581       const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
582       const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
583       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
584 
585       const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
586       const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
587       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
588 
589       const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
590       const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
591       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
592 
593       const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4);
594       const __m512 vk4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 160);
595       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
596 
597       const __m512 vi5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i5);
598       const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192);
599       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
600 
601       const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6);
602       const __m512 vk6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 224);
603       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
604 
605       const __m512 vi7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i7);
606       const __m512 vk7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 256);
607       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
608 
609       const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8);
610       const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288);
611       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
612 
613       const __m512 vi9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i9);
614       const __m512 vk9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 320);
615       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
616 
617       const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10);
618       const __m512 vk10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 352);
619       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
620 
621       const __m512 vi11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i11);
622       const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384);
623       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
624 
625       const __m512 vi12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i12);
626       const __m512 vk12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 416);
627       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
628 
629       const __m512 vi13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i13);
630       const __m512 vk13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 448);
631       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
632 
633       const __m512 vi14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i14);
634       const __m512 vk14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 480);
635       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
636 
637       const __m512 vi15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i15);
638       const __m512 vk15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 512);
639       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
640 
641       const __m512 vi16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i16);
642       const __m512 vk16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 544);
643       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
644 
645       const __m512 vi17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i17);
646       const __m512 vk17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 576);
647       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
648 
649       const __m512 vi18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i18);
650       const __m512 vk18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 608);
651       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
652 
653       const __m512 vi19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i19);
654       const __m512 vk19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 640);
655       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
656 
657       const __m512 vi20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i20);
658       const __m512 vk20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 672);
659       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
660 
661       const __m512 vi21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i21);
662       const __m512 vk21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 704);
663       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
664 
665       const __m512 vi22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i22);
666       const __m512 vk22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 736);
667       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
668 
669       const __m512 vi23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i23);
670       const __m512 vk23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 768);
671       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
672 
673       const __m512 vi24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i24);
674       const __m512 vk24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 800);
675       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
676 
677 
678       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
679       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
680 
681       _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
682       output += c;
683     }
684 
685     output = (float*) ((uintptr_t) output + output_increment);
686   } while (--output_width != 0);
687 }
688